Rooftop Identification and Classification from High-Resolution Aerial Imagery for Photovoltaic Potential Analysis Using Deep Learning

Date

2026-1-29

Author

Ünal, Oğuz Kağan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

1077
views

40
downloads

Cite This

Accurate identification of roof areas and roof types is crucial for assessing solar photovoltaic (PV)potential. This study proposes a two-stage deep learning method for roof analysis using high-resolution aerial RGB images. In the first stage, a semantic segmentation is applied and trained on the Roof Information Dataset 2 (RID2) dataset with a U-Net architecture using the ResNet34 encoder pre-trained on ImageNet. With this trained model, roof areas are detected. In the second stage, the roof areas obtained from the segmentation stage are classified into four different categories (flat, gable/hip, complex, and bugs) using an EfficientNetB0-based classifier. Between these two stages, since the roof polygons are pixel-based, morphological cleaning and contour extraction processes are applied before entering the classification stage, and the resulting polygons are simplified using the Ramer-Douglas-Peucker algorithm. In the RID2 test set, the mean IoU was 0.882 and Dice was 0.9367. Classification accuracy was 80.77% when tested with images on the test set. When the results of the pipeline established in the project are analyzed, it is concluded that the "bugs" class in the classification stage acts as a filter for false positives obtained from the segmentation stage.

Subject Keywords

Roof segmentation, Roof classification, Deep learning, Remote sensing, GIS, Rooftop solar PV potential

URI

https://hdl.handle.net/11511/118369

Collections

Graduate School of Informatics, Term Project